A Study for Automatic Coin Image Recognition Methods

• Main Authors: LIU, FANG-CHIEH, 劉方傑
• Other Authors: Tseng, Chien-Cheng
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/nvy334

碩士 === 國立高雄第一科技大學 === 電腦與通訊工程系碩士班 === 105 === In the field of coin recognition, measurements were traditionally made by physical methods. Recently, computer vision has been widely used in the recognition of modern coins, and also gradually applied to the area of ancient coin recognition. This thesis provides a method, which can effectively recognize modern and ancient coins by computer vision, and obtain high recognition rate with small amount of training data. This thesis will firstly introduce the motivations and backgrounds of coin recognition, as well as the difficulties and challenges of ancient coin recognition. Concerning the problem of translation, rotation, seriously damaged coins and scarcity of training data, this thesis will introduce three algorithms for coins feature vector extraction, including local binary patterns, histogram of oriented gradient, SIFT(scale-invariant feature transform) and its variation－Dense-SIFT, which uses the model of “Bag of visual words” to extract feature and detailed matching. As to the method of feature vector matching, three approaches are studied including histogram matching, earth mover distance and support vector machine. Finally, there are four individual experimental results of single feature vector matching. And, a final recognition result will be decided through a vote made by multiple expert method. Multiple expert method not only increases recognition rate effectively but also decreases the rate of false recognition. Just like the “The Blind Men and the Elephant”, this method can combine the advantages of various feature extraction methods, and obtains the better final result. Eventually, in the experiments of the ancient coin database, it has been shown that the proposed method can overcome various difficulties of coin recognition and achieve a high recognition rate. After learning only small amount of data, the proposed method can be applied to the coin recognition of ancient coins and modern coins.